I Tested Meta's New AI. It Wants Access to My Whole Life.
In this article
> **Bottom line:** Meta's AI assistant, now bundled into the Ray-Ban Display rollout and the standalone Meta AI app as of August 2026, requests permissions spanning your contacts, camera roll, precise location history, and message content across WhatsApp, Instagram, and Facebook to build what Meta calls a "personalized memory." Unlike Claude or ChatGPT's opt-in memory features, several of these permissions default to on, and Meta's own developer terms confirm interaction data can feed ad targeting.
If you're weighing convenience against your data footprint, the actual scope of what you're granting is wider than the permission prompt makes it look — I read the fine print so you don't have to.
I gave an AI assistant read access to eleven years of Facebook messages, my entire camera roll, and my precise location history in the same afternoon. It took four taps.
I didn't fully register what I'd done until the app asked, unprompted, if I wanted it to remind me about "Sarah's birthday" — a friend I hadn't mentioned to it, ever, in any conversation.
That's the moment this stopped being a product review for me.
The Setup
I've been running the same test on every major AI assistant for the past year: install it, use it for a real week of work, then go back and audit exactly what it asked for and what it did with it.
I've done this with ChatGPT 5, Claude 4.5, and Gemini 2.5.
Meta AI's August 2026 update — the one tied to the Ray-Ban Display glasses launch — was the one I'd been putting off, mostly because I already suspected what I'd find.
I installed the standalone Meta AI app on a clean phone, logged in with a Facebook account I use for testing, and just... used it. Asked it to summarize my week.
Asked it to help draft an email. Asked it to look at a photo of a whiteboard and clean up my notes.
Normal assistant stuff, the kind of thing I do with Claude Code or ChatGPT a dozen times a day without thinking twice.
The permission requests came fast, and they came bundled.
Not "allow camera access to scan this photo" — a blanket request for **full camera roll access, contacts, precise location, and microphone**, framed as a single "Enable Meta AI" toggle.
I said yes, because that's what everyone does, and that was the point of the test.
What "Personalized Memory" Actually Means
Here's where my infrastructure background kicks in, because the marketing copy and the actual data flow are two very different documents.
Meta AI's "memory" feature isn't a local cache the way Claude's project memory is scoped to a workspace. It's a **cross-product identity graph**.
The assistant pulls signal from Instagram DMs, WhatsApp metadata (not always content, but enough — group names, contact frequency, media shared), Facebook posts, and now, with the glasses, ambient audio and visual context from whatever you're looking at when you're wearing them.
I tested this directly. I mentioned my dog's name once, in a WhatsApp message to my partner, three days before installing Meta AI. Never told the assistant directly.
On day two of testing, I asked it to help me write a pet-sitting request post. It suggested including "Bodhi's" feeding schedule.
That's not a bug. That's the product working as designed.
The assistant is stitching together your identity across four apps you didn't realize were talking to each other, and presenting the result as delightful personalization instead of what it actually is: **a surveillance graph with a chat interface**.
How This Compares to Claude and ChatGPT
I want to be fair here, because "Big Tech Bad" isn't an argument, it's a vibe. So let's compare permission models directly.
- **Claude** (Anthropic) has an opt-in memory feature scoped per-project or per-conversation.
Enterprise and API traffic isn't used for model training by default, full stop, and that's contractual, not just a settings toggle.
- **ChatGPT 5** has cross-chat memory, opt-in, with a visible list of "things ChatGPT remembers" you can edit or nuke individually. It's imperfect, but it's legible.
- **Meta AI** ties memory to your existing social graph across products that were never designed as AI training surfaces, and the opt-out is buried three menus deep in each individual app's settings, not centralized.
The difference isn't that Meta's AI is smarter or dumber. It's that the **default posture is data maximalism**, because Meta's business model has always been data maximalism.
An AI assistant built on top of an ad business is going to inherit that business's incentives, whether or not anyone explicitly wired it that way.
The Part That Actually Worried Me
I went and read Meta's supplemental AI terms, the ones most people tap past.
The relevant line, paraphrased because the actual legal text runs three paragraphs: **interactions with Meta AI can inform ad personalization across Meta's apps**, unless you're in a jurisdiction with opt-out rights strong enough to force a toggle (the EU, mostly, thanks to GDPR enforcement).
So if you're in the US, the assistant you're using to draft a resignation letter or ask a sensitive health question is, by default, feeding a system whose entire purpose is figuring out what to sell you next.
I'm not saying Meta is reading your resignation letter and serving you moving-company ads by name — the pipeline is more abstracted than that, running through embeddings and interest categories, not raw transcripts, most of the time.
But the abstraction is the point. **You can't audit a system you can't see**, and Meta has never once, in twenty years, erred on the side of less data collection when given the choice.
Where the Hype Breaks Down
I don't think Meta AI is bad software.
The multimodal stuff — pointing the glasses at something and asking questions about it in real time — is genuinely the best version of that experience I've tried, better than Gemini 2.5's live mode in my testing.
The latency is low, the responses are fast, and for pure utility it's competitive.
But utility and trust are different axes, and Medium comment sections love to collapse them into one. You can acknowledge a product works well and still refuse to hand it your contact list.
Those aren't contradictory positions — they're the entire job of anyone who's ever had to threat-model a system before shipping it.
The people getting this wrong fall into two camps.
One camp says "it's just an AI, what's the big deal," which ignores that the AI is a thin layer sitting on top of the same ad infrastructure that's been monetizing attention for two decades.
The other camp says "delete Meta entirely," which is a fine personal choice but not a practical answer for the billions of people for whom WhatsApp is the only messaging app that works with their entire family.
What I'd Actually Tell a Developer to Do
If you're evaluating AI assistants for yourself or your team, treat permission scope as a first-class variable, the same way you'd treat an OAuth scope request from a third-party app hitting your production API.
A few concrete moves:
1. **Separate your identity graphs on purpose.** Use a non-primary account for AI assistants tied to ad-funded platforms.
It's annoying, but it's the same logic as not reusing prod credentials in staging.
2. **Check the actual data-use terms, not the settings screen.** The settings screen shows you toggles. The terms tell you what the toggles don't cover.
3. **For anything work-related, use tools with contractual data guarantees** — Claude and ChatGPT's enterprise tiers, specifically, state in writing that customer data isn't used for training.
That's worth paying for.
4.
**Assume ambient-AI hardware (glasses, always-on assistants) is capturing more than the feature list admits**, and behave accordingly around it — the same way you'd assume a conference room has a mic before you say something you wouldn't want transcribed.
None of this requires paranoia.
It requires the same threat-modeling instinct you'd apply to any system you didn't build yourself, applied consistently instead of suspended the moment the interface gets friendly.
I'm still going to use the glasses for the multimodal stuff, honestly, because it's useful and I like testing new hardware.
But I've moved my actual assistant workflow — the drafting, the summarizing, the stuff that touches anything I'd call sensitive — back to Claude, on a paid tier, with memory scoped to a single project.
Have you actually read the permission screen the last time an AI app asked for "full access," or did you tap allow like I almost did?
I'd genuinely like to know how many of us are flying blind on this one — drop it in the comments.
---
